The proposed research presents a structure of a human-AI collaboration agent to provide dynamic feedback for a Chinese large-unit instruction context. The dataset consists of eight cycles of Chinese large-unit teaching sessions in Grade 5. It includes 320 participants from 32 lessons conducted in 32 classes. Data types involved include lesson plans, dialogue between students and teachers, workbooks, reading notes, vocabulary assignments, short compositions, test scores, graded rubrics for reading and writing, and teacher's feedback notes. The project was designed with Python 3.11, MySQL 8.0, BERT Base Chinese, LightGBM, and Vue teacher dashboard. Learning states of learners can be predicted through features connected with text understanding, task accomplishment, interaction engagement, and progress monitoring. The fixed target labels include stable mastery, partial knowledge, problems with expression, and constant vulnerability. The results prove that the learning state recognition model obtains accuracy of 0.872 ± 0.024, F1 score of 0.851 ± 0.027, and weighted recall of 0.864 ± 0.025. As compared to conventional human-only feedback, the human-AI collaboration improves performance at the following task by +0.184 ± 0.036, reduces mistakes repetitions by −0.217 ± 0.041, and increases unit posttest scores by +6.84 ± 1.27.

Multimodal learning environments are now employing more generative AI for explanations, questions, and feedback. Yet, interface designs are still static while rapid changes are observed in user state and cross-modal needs. This paper proposes an adaptive interface regulation framework to coordinate multimodal presentation by using synchronized text, speech, eye movement, and interaction logs. A controlled experiment was conducted on 96 subjects performing reading comprehension, visual interpretation, and interactive questioning tasks using both static and adaptive interface modes. Logs were annotated according to one unified time scheme, and 84 multimodal features were extracted to classify users' states and regulate interface states. Interaction tension and state-to-interface transition were mathematically represented by two functions. Adaptive interface mode decreased mean time taken for each task from 146.3 ± 18.7 s to 127.7 ± 16.4 s, improved average consistency of responses from 0.742 ± 0.028 to 0.851 ± 0.021 (p < 0.001), and reduced average time for switching to 1.47 ± 0.31 s from 2.31 ± 0.44 s. Best performance was shown for medium- and high-difficulty scenarios.

Legal domain large language models are increasingly used in legal consultation, precedent retrieval, document generation, and AI-assisted adjudication. While these models improve the efficiency of legal information processing, legal hallucinations such as fabricated cases, misquoted statutes, distorted holdings, and broken reasoning chains may undermine the reliability of judicial grounds, the clarity of liability allocation, and procedural legitimacy. To identify hallucination risks in AI-generated legal opinions, this study constructs a dynamic risk identification mechanism based on verifiable legal corpora collected from public statutes, judicial interpretations, judicial documents, and typical cases. The mechanism integrates retrieval-augmented generation, semantic consistency detection, citation validity assessment, and reasoning-chain completeness scoring to classify model outputs into different risk levels. The experimental results show that the proposed model outperforms ordinary prompting, RAG prompting, and self-consistency detection in accuracy, recall, F1-score, and AUC, achieving an overall accuracy of 0.86 and a high-risk recall of 0.89. The mechanism transforms legal hallucination from an opaque generation error into a computable, explainable, and reviewable judicial risk, providing technical support for algorithmic transparency, human oversight, and responsibility allocation in AI-assisted justice.

The parole system is an important part of the criminal punishment execution system in our country. Currently, the risk assessment of parole relies on manual experience, which has problems such as inconsistent standards and insufficient accuracy. Although traditional machine learning can quantify predictions, the single model has limited feature mining and weak generalization ability, making it difficult to meet the needs of judicial practice. Therefore, this paper integrates the advantages of deep learning and ensemble learning, and proposes the LSTM-Adaboost algorithm for predicting the success rate of parole. This algorithm uses LSTM to mine temporal and deep features, and Adaboost to optimize sample weights and improve model accuracy. Compared with various traditional algorithms, LSTM-Adaboost significantly outperforms in terms of accuracy, recall rate, precision, F1 score, AUC, with an accuracy rate of 0.855 and an AUC as high as 0.938, far exceeding the index range of models such as CatBoost and random forest, which are 0.606 to 0.770. It can achieve more accurate and efficient parole risk assessment, providing scientific decision-making support for judicial approval.